fix(protocol): allow 128 chat top_logprobs and forward sampling_logprobs_mode

src/protocols/chat.rs

@@ -264,9 +264,10 @@ pub struct ChatCompletionRequest {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub tools: Option<Vec<Tool>>,
 
-    /// An integer between 0 and 20 specifying the number of most likely tokens to return
+    /// An integer between 0 and 128 specifying the number of most likely tokens to return.
+    /// OpenAI caps this at 20; SGLang accepts up to 128, its PD logprob metadata capacity.
     #[serde(skip_serializing_if = "Option::is_none")]
-    #[validate(range(min = 0, max = 20))]
+    #[validate(range(min = 0, max = 128))]
     pub top_logprobs: Option<u32>,
 
     /// An alternative to sampling with temperature
@@ -364,6 +365,10 @@ pub struct ChatCompletionRequest {
     #[serde(default)]
     pub return_sampling_mask: bool,
 
+    /// Sampling log-prob mode, "selected" or "support" (SGLang extension)
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub sampling_logprobs_mode: Option<String>,
+
     /// Absolute start position for returned routed experts (SGLang extension)
     #[serde(default)]
     pub routed_experts_start_len: i32,

src/protocols/generate.rs

@@ -98,6 +98,10 @@ pub struct GenerateRequest {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub return_sampling_mask: Option<Value>,
 
+    /// Sampling log-prob mode, "selected" or "support"
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub sampling_logprobs_mode: Option<String>,
+
     /// Absolute start position for returned routed experts
     #[serde(default)]
     pub routed_experts_start_len: i32,

tests/routing/test_openai_routing.rs

@@ -618,6 +618,7 @@ async fn test_unsupported_endpoints() {
         return_hidden_states: false,
         return_routed_experts: false,
         return_sampling_mask: None,
+        sampling_logprobs_mode: None,
         return_indexer_topk: false,
         routed_experts_start_len: 0,
         return_prompt_token_ids: false,

tests/spec/chat_completion.rs

@@ -579,6 +579,7 @@ fn test_sglang_extension_fields_roundtrip() {
         "return_hidden_states": true,
         "return_routed_experts": true,
         "return_sampling_mask": true,
+        "sampling_logprobs_mode": "support",
         "routed_experts_start_len": 10,
         "return_cached_tokens_details": true,
         "return_prompt_token_ids": true,
@@ -606,6 +607,7 @@ fn test_sglang_extension_fields_roundtrip() {
     assert!(req.return_hidden_states);
     assert!(req.return_routed_experts);
     assert!(req.return_sampling_mask);
+    assert_eq!(req.sampling_logprobs_mode.as_deref(), Some("support"));
     assert_eq!(req.routed_experts_start_len, 10);
     assert!(req.return_cached_tokens_details);
     assert!(req.return_prompt_token_ids);
@@ -622,6 +624,7 @@ fn test_sglang_extension_fields_roundtrip() {
     assert_eq!(serialized["return_hidden_states"], true);
     assert_eq!(serialized["return_routed_experts"], true);
     assert_eq!(serialized["return_sampling_mask"], true);
+    assert_eq!(serialized["sampling_logprobs_mode"], "support");
     assert_eq!(serialized["routed_experts_start_len"], 10);
     assert_eq!(serialized["return_cached_tokens_details"], true);
     assert_eq!(serialized["return_prompt_token_ids"], true);
@@ -656,6 +659,7 @@ fn test_sglang_extension_fields_default_values() {
     assert!(!req.return_hidden_states);
     assert!(!req.return_routed_experts);
     assert!(!req.return_sampling_mask);
+    assert!(req.sampling_logprobs_mode.is_none());
     assert_eq!(req.routed_experts_start_len, 0);
     assert!(!req.return_cached_tokens_details);
     assert!(!req.return_prompt_token_ids);
@@ -695,6 +699,7 @@ fn test_sglang_optional_fields_omitted_when_none() {
         "max_dynamic_patch", "min_dynamic_patch", "rid", "extra_key", "cache_salt",
         "priority", "bootstrap_host", "bootstrap_port", "bootstrap_room",
         "routed_dp_rank", "disagg_prefill_dp_rank", "data_parallel_rank",
+        "sampling_logprobs_mode",
     ];
     for field in omitted {
         assert!(
@@ -704,3 +709,18 @@ fn test_sglang_optional_fields_omitted_when_none() {
         );
     }
 }
+
+#[test]
+fn test_top_logprobs_accepts_up_to_128() {
+    for (count, valid) in [(20, true), (21, true), (128, true), (129, false)] {
+        let req: ChatCompletionRequest = serde_json::from_value(json!({
+            "model": "test-model",
+            "messages": [{"role": "user", "content": "hello"}],
+            "logprobs": true,
+            "top_logprobs": count
+        }))
+        .expect("should deserialize");
+        assert_eq!(req.validate().is_ok(), valid, "top_logprobs={count}");
+        assert_eq!(serde_json::to_value(&req).unwrap()["top_logprobs"], count);
+    }
+}

tests/spec/generate.rs

@@ -8,6 +8,7 @@ fn test_generate_extension_fields_roundtrip() {
         "lora_backfill_paths": {"adapter@1": "/shared/adapter"},
         "return_hidden_states": true,
         "return_routed_experts": true,
+        "sampling_logprobs_mode": "support",
         "routed_experts_start_len": 10,
         "return_prompt_token_ids": true,
         "require_reasoning": true,
@@ -24,6 +25,7 @@ fn test_generate_extension_fields_roundtrip() {
         serde_json::from_value(json_with_extensions).expect("should deserialize");
     assert!(req.return_hidden_states);
     assert!(req.return_routed_experts);
+    assert_eq!(req.sampling_logprobs_mode.as_deref(), Some("support"));
     assert_eq!(req.routed_experts_start_len, 10);
     assert!(req.return_prompt_token_ids);
     assert!(req.require_reasoning);
@@ -41,6 +43,7 @@ fn test_generate_extension_fields_roundtrip() {
     );
     assert_eq!(serialized["return_hidden_states"], true);
     assert_eq!(serialized["return_routed_experts"], true);
+    assert_eq!(serialized["sampling_logprobs_mode"], "support");
     assert_eq!(serialized["routed_experts_start_len"], 10);
     assert_eq!(serialized["return_prompt_token_ids"], true);
     assert_eq!(serialized["require_reasoning"], true);
@@ -62,6 +65,7 @@ fn test_generate_extension_fields_defaults() {
     let req: GenerateRequest = serde_json::from_value(json_minimal).expect("should deserialize");
     assert!(!req.return_hidden_states);
     assert!(!req.return_routed_experts);
+    assert!(req.sampling_logprobs_mode.is_none());
     assert_eq!(req.routed_experts_start_len, 0);
     assert!(!req.return_prompt_token_ids);
     assert!(!req.require_reasoning);
@@ -91,6 +95,7 @@ fn test_generate_optional_fields_omitted_when_none() {
         "max_dynamic_patch",
         "min_dynamic_patch",
         "lora_backfill_paths",
+        "sampling_logprobs_mode",
     ];
     for field in omitted {
         assert!(